Pairwise Surrogate Learning for Neural Architecture Ranking in Digital Pathology
11th International Conference on Machine Learning Technologies, ICMLT 2026, Berlin, Almanya, 20 - 22 Mayıs 2026, ss.650-654, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/icmlt69916.2026.11688928
- Basıldığı Şehir: Berlin
- Basıldığı Ülke: Almanya
- Sayfa Sayıları: ss.650-654
- Anahtar Kelimeler: digital pathology, NAS, surrogate-based learning
- Marmara Üniversitesi Adresli: Evet
Özet
Neural Architecture Search can automate the design of accurate and efficient deep networks for whole-slide image analysis in digital pathology, but exhaustive evaluation of candidate architectures is computationally expensive. This study proposes a pairwise surrogate-learning framework to predict relative performance between architectures generated from two NAS formulations: Non-Plato NAS and Path-Based NAS. For each of four pathology datasets (EBHI, BCNB, TCGA, and SPIDER), 200 architectures per search space are sampled via Random Search and trained to obtain F1-based fitness. Ordered architecture pairs are then labeled by which model performs better, enabling supervised learning of auxiliary predictors. We compare tree-based models and graph neural networks (GNNs) using both pairwise classification metrics and ranking/selection metrics after aggregating pairwise outcomes into global rankings. Results show boosted trees are most robust, while GNNs benefit from structured path-based encodings.